Oh. No, I did not!
Welp, that fixed it. Thank you. Updated code, internal notes-to-self and all:
class ImageScalarDataset(ImageDataset):
def __init__(self, df:DataFrame, path_column:str='file_path', dependent_variable:str=None):
# The superclass does nice things for us like tensorizing the numpy
# input
super().__init__(df[path_column], np.array(df[dependent_variable], dtype=np.float32))
# Old FastAI uses loss_fn, new FastAI uses loss_func
self.loss_func = layers.MSELossFlat()
self.loss_fn = self.loss_func
# We have only one "class" (i.e., the single output scalar)
self.classes = [0]
def __len__(self)->int:
return len(self.y)
def __getitem__(self, i):
# return x, y | where x is an image, and y is the scalar
return open_image(self.x[i]), self.y[i]
1 Like
Hi - I have a similar use case to yours. I want to classify the age of a photo and have processed the data to have the age number as the filename. My approach was to build the image data bunch and just use a different loss function (MSELoss rather than Cross Entropy) and hoped that would do it but ran into issues:
data = (ImageFileList.from_folder(path_img)
.label_from_func(get_float_labels)
.random_split_by_pct(valid_pct=0.2)
.datasets()
.transform(get_transforms(), size=224)
.databunch(bs=bs).normalize(imagenet_stats)
)
My classes look like this
class MSELossFlat2(nn.MSELoss):
"Same as `nn.MSELoss`, but flattens input and target."
def forward(self, input:Tensor, target:Tensor) -> Rank0Tensor:
return super().forward(input.view(-1).float(), target.view(-1).float() )
learn = create_cnn(data, models.resnet34)
learn.loss_func = MSELossFlat2()
learn.fit_one_cycle(4)
If I don’t modify my MSELoss then I get an error about expecting a Float and getting a long form target. If i do this I get a mismatch error:
RuntimeError: input and target shapes do not match: input [7040], target [64] at /opt/conda/conda-bld/pytorch-nightly_1540036376816/work/aten/src/THCUNN/generic/MSECriterion.cu:12
Any tips would be appreciated. Thanks!
Did anyone work on this recently? I want to predict a scalar with an image too.
I have a dataframe with the filenames in one column and the scalar I want to predict on another column.
I can’t get the custom dataset to work. Where does the ImageDataset class come from?
Is there any other/easier way to get this to work in fastai v1 ? In fastai 0.7 you just had to change one argument to the ImageClassifierData factory method
EDIT:
nvm, I was just to stupid to save my scalar as a float instead of an integer. Now the last layer has one output like expected. It seems to work with the standard ImageDataBunch.from__df()
Would you be willing to provide an example of predicting scalars from images using the standard ImageDataBunch.from_df() ?
I have a very similar problem. I have a data frame with two columns, the first “filename” and the second “age”, saved as a string object and int64 respectively.
When I run the following code:
data = ImageDataBunch.from_df(path=path, df=df, size=224, bs=64)
arch = models.resnet34
learn = create_cnn(data, arch, metrics=MSELossFlat)
lr = 5e-2
learn.fit_one_cycle(1, slice(lr))
I get the same error message as jamesp:
RuntimeError: bool value of Tensor with more than one value is ambiguous
EDIT: I think I figured it out. This blog post/code are very helpful:
The key is to use ‘label_cls=FloatList’ when constructing the labels. This tells fastai to expect a regression problem, e.g:
data = (ImageItemList.from_csv(path, ‘csv_file’, folder=’’, suffix=’’)
.random_split_by_pct(0.2)
.label_from_df(cols=1, label_cls=FloatList)
.transform(tfms, size=224)
.databunch())
data.normalize(imagenet_stats)
1 Like
Could anyone provide some insight on how this might be extended to predict multiple scalar outputs? I’m having trouble figuring out a way to do this.
Another post with the same question that is unresolved.
If you pass a list instead of 1 to cols, it should work.
Not working on my problem. I needed to change a lot of code and I think now it is not working for other problems. I will send code here if everything worked well.
Should be fixed in master now.
I’m acctualy using this for text data so I’m not sure is this something related to that but I got this error when I tried to train the model.
RuntimeError: Expected object of scalar type Long but got scalar type Float for argument #2 'other'
targs: tensor([[0., 0., 1.],
[0., 1., 0.],
[1., 0., 0.],
[1., 0., 0.],
[1., 0., 0.],
[0., 1., 0.]], device=‘cuda:0’)
input: tensor([[2],
[2],
[0],
[0],
[0],
[0]], device=‘cuda:0’)
Targets can be anything from 0 to 1 and I need three different values.
data = (TextList.from_csv('.', text_data_with_three_targets.csv', vocab=data.vocab)
.random_split_by_pct()
.label_from_df(cols=[1,2,3],label_cls=FloatList)
.databunch(bs=32,no_check=True))When I try that, I run into the following error when I try to create my DataBunch:
TypeError: can't convert np.ndarray of type numpy.object_. The only supported types are: double, float, float16, int64, int32, and uint8.
This occurs when the summary (repr) of the DataBunch is about to be outputted, specifically in tensor() method, where there still seems to be an issue as referred to in this line:
# XXX: Pytorch bug in dataloader using num_workers>0; TODO: create repro and report
How can I get around this?
It’s hard to help without the rest of your code and the full error message.
I will provide those tomorrow and I will also try to debug it by myself.
You don’t have the latest code, so it’s normal it doesn’t work.
path = Path('/path/to/texts')
data = (TextList.from_folder(path,extensions='.txt')
.random_split_by_pct(0.1)
.label_for_lm()
.databunch(bs=32))
learn = language_model_learner(data, pretrained_model=URLs.WT103_1, drop_mult=0.3)
learn.load(Path('/path/to/fine_tuned/language_model'))
data2 = (TextList.from_csv('.', 'example.csv', vocab=data.vocab)
.random_split_by_pct()
.label_from_df(cols=[1,2,3],label_cls=FloatList)
.databunch(bs=32))
learn = text_classifier_learner(data2,drop_mult=0.5)
learn.load_encoder('/path/to/fine_tuned_enc')
learn.freeze()
learn.fit_one_cycle(1, 2e-2, moms=(0.8,0.7))
Error message:
RuntimeError Traceback (most recent call last)
in ()
----> 1 learn.fit_one_cycle(1, 2e-2, moms=(0.8,0.7))
~/Downloads/fastai-master (1)/fastai-master/fastai/train.py in fit_one_cycle(learn, cyc_len, max_lr, moms, div_factor, pct_start, wd, callbacks, **kwargs)
20 callbacks.append(OneCycleScheduler(learn, max_lr, moms=moms, div_factor=div_factor,
21 pct_start=pct_start, **kwargs))
---> 22 learn.fit(cyc_len, max_lr, wd=wd, callbacks=callbacks)
23
24 def lr_find(learn:Learner, start_lr:Floats=1e-7, end_lr:Floats=10, num_it:int=100, stop_div:bool=True, **kwargs:Any):
~/Downloads/fastai-master (1)/fastai-master/fastai/basic_train.py in fit(self, epochs, lr, wd, callbacks)
171 callbacks = [cb(self) for cb in self.callback_fns] + listify(callbacks)
172 fit(epochs, self.model, self.loss_func, opt=self.opt, data=self.data, metrics=self.metrics,
--> 173 callbacks=self.callbacks+callbacks)
174
175 def create_opt(self, lr:Floats, wd:Floats=0.)->None:
~/Downloads/fastai-master (1)/fastai-master/fastai/basic_train.py in fit(epochs, model, loss_func, opt, data, callbacks, metrics)
93 except Exception as e:
94 exception = e
---> 95 raise e
96 finally: cb_handler.on_train_end(exception)
97
~/Downloads/fastai-master (1)/fastai-master/fastai/basic_train.py in fit(epochs, model, loss_func, opt, data, callbacks, metrics)
88 if not data.empty_val:
89 val_loss = validate(model, data.valid_dl, loss_func=loss_func,
---> 90 cb_handler=cb_handler, pbar=pbar)
91 else: val_loss=None
92 if cb_handler.on_epoch_end(val_loss): break
~/Downloads/fastai-master (1)/fastai-master/fastai/basic_train.py in validate(model, dl, loss_func, cb_handler, pbar, average, n_batch)
53 if not is_listy(yb): yb = [yb]
54 nums.append(yb[0].shape[0])
---> 55 if cb_handler and cb_handler.on_batch_end(val_losses[-1]): break
56 if n_batch and (len(nums)>=n_batch): break
57 nums = np.array(nums, dtype=np.float32)
~/Downloads/fastai-master (1)/fastai-master/fastai/callback.py in on_batch_end(self, loss)
248 "Handle end of processing one batch with `loss`."
249 self.state_dict['last_loss'] = loss
--> 250 stop = np.any(self('batch_end', not self.state_dict['train']))
251 if self.state_dict['train']:
252 self.state_dict['iteration'] += 1
~/Downloads/fastai-master (1)/fastai-master/fastai/callback.py in __call__(self, cb_name, call_mets, **kwargs)
196 def __call__(self, cb_name, call_mets=True, **kwargs)->None:
197 "Call through to all of the `CallbakHandler` functions."
--> 198 if call_mets: [getattr(met, f'on_{cb_name}')(**self.state_dict, **kwargs) for met in self.metrics]
199 return [getattr(cb, f'on_{cb_name}')(**self.state_dict, **kwargs) for cb in self.callbacks]
200
~/Downloads/fastai-master (1)/fastai-master/fastai/callback.py in <listcomp>(.0)
196 def __call__(self, cb_name, call_mets=True, **kwargs)->None:
197 "Call through to all of the `CallbakHandler` functions."
--> 198 if call_mets: [getattr(met, f'on_{cb_name}')(**self.state_dict, **kwargs) for met in self.metrics]
199 return [getattr(cb, f'on_{cb_name}')(**self.state_dict, **kwargs) for cb in self.callbacks]
200
~/Downloads/fastai-master (1)/fastai-master/fastai/callback.py in on_batch_end(self, last_output, last_target, **kwargs)
283 if not is_listy(last_target): last_target=[last_target]
284 self.count += last_target[0].size(0)
--> 285 self.val += last_target[0].size(0) * self.func(last_output, *last_target).detach().cpu()
286
287 def on_epoch_end(self, **kwargs):
~/Downloads/fastai-master (1)/fastai-master/fastai/metrics.py in accuracy(input, targs)
29 print('targs:',targs)
30 print('input:',input)
---> 31 return (input==targs).float().mean()
32
33 def accuracy_thresh(y_pred:Tensor, y_true:Tensor, thresh:float=0.5, sigmoid:bool=True)->Rank0Tensor:
RuntimeError: Expected object of scalar type Long but got scalar type Float for argument #2 'other'
example.csv
Ah, this is because of the metrics always set to accuracy. You can remove it by passing metrics=[] in your learner creation. I’ll fix it this morning.
1 Like
Thank you @sgugger ! It is working now.
Do you think there might be idea to create own accuracy for this kind of multi output problem? It could calculate the accuracy for every pair and then calculate the mean of those.
Coming back to this after the v1 API seems to have stabilized, I see ImageDataBunch and ImageItemList. It seems that ImageDataBunch's initializers are only for classifiers (since I don’t see a way to set label_cls), so if we want to build a regressor without writing a custom initializer, we need to use ImageItemList. Is that accurate?
ImageDataBunch and its factory methods are for beginners. You can do a regression problem with it if your targets are exactly like the library would expect them to be (floats) but you should really use the data block API (and ImageItemList) to have the maximum flexibility.
1 Like